Enflame CloudBlaze S60
Enflame's third-generation AI inference accelerator, based on the self-developed GCU320 (Suiyuan 320) chip, released in March 2024, with a maximum power of 300W, targeting large-scale data center deployment and large-model inference (full-spec DeepSeek all-in-one appliance solutions).
Core Specifications
| Specification | Value |
|---|---|
| Architecture | GCU320 (Suiyuan 320) |
| Process | Not disclosed |
| TDP | 300 W (official product manual, maximum power) |
| Memory | 48 GB (official product manual) |
| Memory Bandwidth | 672 GB/s (official product manual) |
| Precision Support | FP32 / FP16 / BF16 / INT8 (peak per-precision compute not published by the vendor) |
| Interface | PCIe Gen5 x16, full-height full-length dual-slot |
| Video Decode | Up to 256 channels |
| ECC / Security | ECC, Secure Boot, SR-IOV (4 VF) |
| Release | 2024-03 |
| Price | Not disclosed |
📌 Data correction (2026-09 cross-validation): This page previously recorded "1.6 TB/s bandwidth / FP16 100 TFLOPS (estimated)", which does not match the official manual. Enflame's official S60 product manual confirms: 48GB memory, 672 GB/s, PCIe Gen5 x16, maximum power 300W, support for FP32/FP16/BF16/INT8, but does not publish peak per-precision compute — corrected per the "no estimates when undisclosed" principle.
Technical Highlights
- Third-generation inference card: Based on the Suiyuan 320 (GCU320) chip, Enflame's third-generation AI inference product
- Large-model inference optimization: Optimized for large language model inference scenarios such as LLaMA and DeepSeek; the vendor offers an all-in-one appliance supporting the full 671B model
- Search/ads/recommendation support: Sustains tens of billions of daily calls for Tencent's search, ads, and recommendation workloads (per the IPO prospectus)
- Easy migration: Broad model coverage and strong usability, supporting smooth migration from NVIDIA GPUs
- High-density deployment: Maximum power 300W with passive air cooling, suitable for large-scale data center deployment
Product Positioning
The CloudBlaze S60 is Enflame's new-generation AI inference accelerator for large-scale data center deployment, targeting the NVIDIA L4/L40. As a third-generation product, the S60 significantly improves on memory capacity (48GB) and video decoding (256 channels) over the previous-generation CloudBlaze i20 (16GB), and is a representative product of Enflame's "card–model–system" closed loop (full DeepSeek model adaptation).
Application Scenarios
- Large language model inference (LLaMA, DeepSeek, ChatGLM, etc.)
- Search, advertising, and recommendation system inference
- Computer vision inference (CV, 256-channel video decode)
- Natural language processing inference (NLP)
- Large-scale data center inference deployment
Reference Price
- Official pricing: Not disclosed
Related Products
- CloudBlaze T20 — Second-generation training card
- CloudBlaze T21 — Second-generation training card (OAM)
- CloudBlaze i20 — Second-generation inference card
- Enflame L600 — Fourth-generation unified training/inference chip